Nonconvex Optimization Using a Fokker-planck Learning Machine
نویسندگان
چکیده
A new algorithm for nonconvex optimization by means of a so-called Fokker-Planck Learning Machine is proposed in this paper. This is done by considering the Fokker-Planck (FP) equation related to continuous simulated annealing, which has been proven to convergence to the global optimum under certain conditions. An approximate solution to the FP equation is sought by parametrizing the transition density by means of Gaussian sum approximations (Radial Basis Function networks). Like in genetic algorithms a population of points is considered. At each generation the points are generated based upon the transition density and the density is updated according to the FP equation.
منابع مشابه
The Fokker-planck Machine as Deterministic Ensemble Annealing 1 the Fokker-planck Machine as Deterministic Ensemble Annealing
In this paper we discuss the relation between ensemble learning, deterministic annealing and Markov chain sample and optimization methods on the one hand and the Fokker-Planck Machine on the other hand. The latter was introduced as an alternative method for high-dimensional non-convex optimization, as often encountered in training of neural networks. We demonstrate how the Fokker-Planck Machine...
متن کاملBayesian Learning and the Fokker-planck Machine 1 Bayesian Learning and the Fokker-planck Machine
In this paper we discuss the links between simulated annealing, Langevin optimisation, Markov Chain Monte Carlo methods and ensemble learning on the one hand and the Fokker-Planck machine on the other hand. The FP Machine was introduced as an alternative method for global optimization. Where most optimisation methods follow point-trajectories through search space (either deterministic or stocha...
متن کاملNumerical Studies and Simulation of the Lower Hybrid Waves Current Drive by using Fokker – Planck Equation in NSST and HT-7 Tokamaks
Recent experiments on the spherical tokamak have discovered the conditions to create a powerful plasma and ensure easy shaping and amplification of stability, high bootstrap current and confinement energy. The spherical tours (ST) fusion energy development path is complementary to the tokamak burning plasma experiment such as NSTX and higher toroidal beta regimes and improves the design of a po...
متن کاملCharacterization of the convergence of stationary Fokker-Planck learning
The convergence properties of the stationary Fokker-Planck algorithm for the estimation of the asymptotic density of stochastic search processes is studied. Theoretical and empirical arguments for the characterization of convergence of the estimation in the case of separable and nonseparable nonlinear optimization problems are given. Some implications of the convergence of stationary Fokker-Pla...
متن کامل